BishnupriyaManipuri/nllb-bpy-beng-v8-5-3-merged

NLLB Bishnupriya Manipuri V8.5.3 - Merged

Status: ✅ Production - Inference API Ready

This is the merged version of V8.5.3 with LoRA weights integrated into NLLB-200-distilled-600M. Use this repo for Hugging Face Inference API and Inference Endpoints.

Base model: facebook/nllb-200-distilled-600M

Adapter: Emarthar/nllb-bpy-beng-v8-5-3

Merge method: PEFT mergeandunload()

V8.5.3 Fixes:

• —1000x weighted number+noun patterns
• —Fixes repetition bug: য়াংখেইহান লেরিক (not লেরিকহান লেরিকহান)
• —Uses benBeng token instead of asmBeng

Test Results:

• —Fifty books → য়াংখেইহান লেরিক
• —Twenty books → আককুরিহান লেরিক
• —My father works → মর বাবা কাম করের
• —The sun is hot → বেলীগ তপ্তা ইসে

Usage:

from transformers import AutoModelForSeq2SeqLM, AutoTokenizer

model = AutoModelForSeq2SeqLM.from_pretrained('BishnupriyaManipuri/nllb-bpy-beng-v8-5-3-merged')
tokenizer = AutoTokenizer.from_pretrained('BishnupriyaManipuri/nllb-bpy-beng-v8-5-3-merged')

def translate(text):
 tokenizer.src_lang = 'eng_Latn'
 inputs = tokenizer(text, return_tensors='pt')
 out = model.generate(**inputs, forced_bos_token_id=tokenizer.convert_tokens_to_ids('ben_Beng'))
 return tokenizer.decode(out[0], skip_special_tokens=True)

Training data: See training_data.csv in this repo

Original adapter: Emarthar/nllb-bpy-beng-v8-5-3

License: MIT - Free for commercial use